Compressed Deep Learning Model For Human Activity Recognition

Researchers are becoming more interested in Human Activity Recognition (HAR) because wide varity application of HAR.Due to the fact that wearable sensors such as accelerometers and gyroscopes may give time-series data for the analysis of human behavior, there has been a spike in interest in Human Activity Recognition (HAR). Effective HAR necessitates the extraction of critical temporal aspects from the data. Conventional methods are time-consuming, frequently necessitate expert knowledge, and involve extensive feature engineering and data processing. A breakthrough in automated feature detection was made possible by a unique Deep Neural Network that combined Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN). On the WISDM dataset, this model’s accuracy was 99.74 percent. Its huge size, however, makes integration onto microcontrollers difficult. Techniques like quantization and pruning are used to address this, which makes these models appropriate for edge devices.